05-05-2011, 10:20 AM
Abstract
An overview is given of a novel vision system for locating,recognising and tracking multiple vehicles, using a singlemonocular camera mounted on a moving vehicle1. 3-D model-based techniques are employed to obtain detailedinformation about vehicle movements. Egomotion estimationis performed by inter-frame tracking of features on theground plane. Vehicle detection and hypothesis generationis performed using a template correlation technique. Oncedetected and identified, vehicles are tracked using dynamicfiltering. The resulting information about 3-D vehicle motionsis passed to a collision alert system which can warnthe driver of potential hazards. The overall approach isdemonstrated on a typical motorway image sequence
.Keywords: Model-based vision, surveillance systems,traffic scene analysis, vehicle tracking.
1. Introduction
In today’s crowded traffic conditions the cost of an accidentcan be heavy both in human terms and in its effect ona road network working at or near to full capacity. Any reductionin the number and severity of accidents would havelarge social and economic benefits. Recent work in computervision may provide the means for such a reduction.With the rapidly falling costs of cameras and computers itis now feasible to equip a car with an autonomous artificialvision system which can monitor nearby vehicles, keeptrack of their motions and warn the driver whenever there isdanger of a collision.The model-based tracking system developed at Reading[15, 16] can locate and reason about the positions of 3-D vehicles in a known 3-D world. Model-based algorithmsare robust to variations in illumination and colour, and tochanges in the position of the object relative to the camera.They deal properly with occlusions and make it possibleto impose object-centered dynamic constraints. The algorithmscan be easily extended to deal with camera movementand to utilise the information from several cameras.The model gives the tracker a level of stability and accuracythat is impossible to achieve in purely bottom-up methodsof tracking.This paper reports new results on the application ofmodel-based vision to collision alert, wih the ultimate aimof producing a complete integrated closed-loop system forvehicle detection, hypothesis generation and verification,tracking, trajectory analysis and collision alert. The mainsource of information for the system is a sequence of imagestaken by a single video camera.The first part of this paper describesmethods for estimatingthe camera motion (egomotion) from point matches andfor detecting far-off target vehicles using template matching.In the second part, model-based techniques are employedto infer 3-D trajectories of vehicles from an imagesequence. Applications to collision alert are discussed.
1.1. Background work
Machine vision has huge potential for the detection andidentification of road vehicles, and for monitoring theirpositions and interactions over time. To date, traffic visionsystems have largely assumed static cameras (e.g.[3, 10, 15, 16]). In contrast, this work uses image sequencestaken by a cameramounted in amoving vehicle. The IntelligentVehicle Symposia (esp. 1994-1996) describe a numberof vision systems for detecting and tracking road lanes, vehiclesand obstacles using monocular, binocular or trinocularvision. Existing methods for motion estimation andtracking typically employ optical flow or other 2-D imagebased techniques (e.g. [1, 2, 7, 11, 14]). These approachesare adequate for tasks such as the estimation of time to collisionor object detection but i) they cope poorly with overlappingvehicles, shadows, rain etc; and ii) in the absenceof a model they are not accurate enough to generate an understandingof the 3-D form and position of vehicles. Themodel-based approach can overcome the difficulties i and iieven if the camera is moving.To our knowledge this is the first time that a full 3-Dmodel-based vision system has been employed for monitoringroad traffic from a car-mounted camera. The closestrelated work is [4] in which model-based tracking of lanesand objects is performed, but our work goes further in thatit includes full 3-D models and addresses collision alert issues.
Download full report
http://wjm.tyai.tyc.edu.tw/~jmwang/paper...hicles.pdf